Jul 2026· e+i Elektrotechnik und Informationstechnik· Vol 143, pp. 476 - 487· 0 citations· 44 references
TL;DR
A systematic comparison of Long Short-Term Memory and transformer-based architectures for deterministic short-term PV power forecasting using publicly accessible data from multiple climatic regions highlights the advantages of attention-based sequence modeling for PV applications and offers practical guidance on feature design, input horizon selection, and hyperparameter ranges for future data-driven PV forecasting studies.
Abstract
Accurate short-term photovoltaic (PV) power forecasting is increasingly important for maximizing on-site self-consumption and ensuring reliable grid integration as decentralized PV deployment grows. This paper presents a systematic comparison of Long Short-Term Memory (LSTM) and transformer-based architectures for deterministic short-term PV power forecasting using publicly accessible data from multiple climatic regions. The dataset combines multi-year 15-minute PV power measurements from nine plants with corresponding meteorological variables, and it is followed by a unified preprocessing pipeline that includes outlier treatment, interpolation, feature scaling, and correlation-based feature selection. Several feature subsets and input window lengths are evaluated, and Bayesian hyperparameter optimization is employed to refine model configurations for both architectures. The results indicate that using all meteorological variables except cloud coverage with a 2-day input window yields the best performance. Under this configuration, the transformer model outperforms the LSTM model, achieving on average test errors of MSE = 0.0038 and MAE = 0.0265, compared to 0.0055 and 0.0343 for the LSTM, respectively. An analysis of time-resolved residuals shows that both models exhibit the largest errors around noon, while the transformer provides a consistently narrower error distribution over the diurnal cycle. These findings highlight the advantages of attention-based sequence modeling for PV applications and offer practical guidance on feature design, input horizon selection, and hyperparameter ranges for future data-driven PV forecasting studies.
With the global shift toward green energy, solar photovoltaic (PV) power has expanded rapidly. However, the unpredictable nature of PV generation, caused by changing weather conditions like irradiance and temperature, challenges grid stability and power scheduling. Therefore, developing smart forecasting models for high-precision PV power prediction is essential for modern grid management. This paper introduces an optimized forecasting framework using Long Short-Term Memory (LSTM) networks. By integrating historical power generation data with localized meteorological factors, a multivariate predictive model was developed and validated using empirical data from a 100 MW PV plant. Based on Pearson correlation analysis, four key features—temperature, direct normal irradiance (DNI), relative humidity, and cloud cover—were chosen as the main drivers of PV output. A multivariate LSTM model was then trained and carefully tested using time series cross-validation. Results show the multi-feature architecture consistently outperforms and surpasses single-feature benchmarks. Specifically, the model achieved a peak R2 of 0.9864 and minimum MAE of 1.4057 kW. Across four validation sets, R2 remained stable (0.9755–0.9836), with most errors tightly bounded within ±5 kW. These findings confirm the model’s excellent accuracy and its value for improving power system dispatch and resource planning.
Rooftop photovoltaic (PV) systems exhibit significant variability in short-term electricity generation due to differences in panel tilt and azimuth angles. Neglecting this geometric diversity may reduce forecasting accuracy, particularly in distribution networks with heterogeneous rooftop PV installations. This study investigates short-term rooftop PV forecasting using an orientation-aware modeling framework based on a real-scale experimental dataset. Five 300 W monocrystalline PV panels were monitored hourly for approximately 1 year under four different tilt angles and seven azimuth angles, representing 28 distinct orientation configurations. Temporal variables (the cosine of day and hour), meteorological variables, and the cosine of the solar incidence angle were used as input features. Three forecasting model structures (Models I–III), each defined by a different set of input features, were developed. Each model was tested using the Persistence Method (PM), Long Short-Term Memory (LSTM), and Gradient Boosted Regression Trees (GBRT). Among the evaluated models, Model II, which included the cosine of the solar incidence angle, achieved the best overall performance. In this model, GBRT outperformed both LSTM and PM by yielding the lowest mean test root mean square error (RMSE) (13.56 ± 2.18 Wh), the highest mean test R
2
(0.973 ± 0.009), and the lowest mean test mean absolute percentage error (MAPE) (18.04 ± 3.65). The results demonstrate that incorporating physically meaningful orientation-aware features substantially improves forecasting accuracy for heterogeneous rooftop PV systems. These findings highlight the importance of panel orientation in hour-ahead rooftop PV forecasting. Consequently, the proposed approach can improve the accuracy of distributed PV generation forecasts and net demand forecasts at the distribution level compared to traditional aggregation-based approaches relying on general solar radiation data.
H. Cevik, Mustafa Arslan, Mehmet Çunkaş· PeerJ Computer Science· 0 citations
The large-scale integration of photovoltaic (PV) systems into modern power grids impacts operational challenges in reducing the intermittency of solar irradiance. Short-term forecasting, especially with accurate result is essential for grid stability, economic dispatch, and demand-side management. However, the development of robust deep-learning models is frequently constrained by the limited readiness of high-quality and labeled actual generation data. This paper presents a simulation-driven methodology to address this data scarcity problem. A synthetic PV generation dataset is created using a deterministic mathematical model. This method captures the diurnal solar cycle, augmented with Gaussian stochastic noise to simulate the effects of cloud cover and atmospheric turbulence. A Long Short-Term Memory (LSTM) network in two layer is trained on the synthetic dataset and subsequently validated against real hourly PV generation data for Bali, Indonesia. This data was obtained from the PVGIS-ERA5 database. The method performs 1-hour-ahead sequence-to-point forecasting using a 6-hour retrospective window. A persistence baseline model is employed for comparative benchmarking. This experiment found that the LSTM outperforms the persistence baseline on both datasets significantly. For the empirical Bali PV dataset, the LSTM attains a Mean Absolute Error (MAE) and a Root Mean Square Error (RMSE) of 29.75 W and 43.31 W, respectively, with a corresponding R2 value of 0.9474. Compared to the persistence benchmark's MAE of 69.00 W, the network successfully cuts prediction error by 56.9%. These results validate both the simulation-driven training approach and the LSTM's capability for short-term solar forecasting under tropical conditions.
I. D. Saputra, Nicola Schulz, I Nyoman Kusuma Wardana et al.· 2026 International Conferenc...· 0 citations
Accurate short-term rooftop photovoltaic (PV) power forecasting is important for energy system management because forecasting errors directly affect local scheduling, reserve coordination, and distributed PV balancing under rapidly changing meteorological conditions. This paper proposes a CMDHOLE-Transformer-LSTM framework for deterministic one-step-ahead forecasting of aggregated rooftop PV power in a campus-level distributed PV setting. The model combines a Transformer module for global dependency extraction, a long short-term memory (LSTM) module for temporal-memory refinement, and a Cauchy-mutation-improved DHOLE (CMDHOLE) algorithm for hyperparameter optimization. Experiments are conducted on the public Hong Kong University of Science and Technology rooftop PV dataset, in which outputs from multiple rooftop PV stations are aggregated into a single campus-level PV power series. Historical PV power and eight meteorological variables are used as inputs under a chronological 80%/10%/10% training-validation-testing partition. Because nighttime and near-zero PV samples are retained, percentage-based metrics are excluded, and performance is evaluated using mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R2). The proposed model achieves a test-set MAE of 25.4631 kW, an RMSE of 57.8833 kW, and an R2 of 0.96657. Relative to the strongest non-proposed baseline, Extreme Gradient Boosting, MAE and RMSE are reduced by 7.37% and 3.45%, respectively. The ablation and optimizer-comparison results further show that both the hybrid Transformer-LSTM backbone and the CMDHOLE optimization strategy contribute meaningfully to the final forecasting performance. Overall, the proposed framework provides an effective solution for short-horizon rooftop PV forecasting in campus-level distributed energy-management applications.
Rui Tian, Gang Liu· 2026 5th International Confe...· 0 citations
This study provides an in-depth comparative analysis of four state-of-the-art neural architectures, confirming that high-fidelity point forecasts and rigorously quantified uncertainty can be achieved simultaneously, providing a clear path toward more dependable PV dispatch, reserve allocation, and market participation.
Saloni Dhingra, G. Gruosso, G. Storti Gajani· Neural computing & applicati...· 0 citations
A hybrid PV forecasting framework that combines stacking ensemble learning with a targeted residual correction strategy, and demonstrates that analyzing error distribution and forecasting robustness provides valuable insights beyond conventional aggregate metrics, contributing to the development of more reliable photovoltaic forecasting systems.
Khawla Oufrit, A. Mouadili, M. Zazoui· EPJ Web of Conferences· 0 citations